Cherry Agent- Autonomous Finance Ops

Cherry Agent Architecture
Invoce & receipt Ingestion
Autonomous Reconciliation workflow
Human approval & Exception handling
Auditing Completed

Inspiration

Small businesses spend far too much time doing work that their finance software already has enough information to help complete.

A typical supplier transaction can involve opening an email, downloading an invoice, extracting the amount and VAT, choosing an accounting category, finding the corresponding bank transaction, reconciling it, requesting approval, attaching evidence and repeating the process again at month-end.

Most accounting software still turns this into a long to-do list.

We wanted to ask a different question:

What if finance software could actually finish the work?

That inspired Cherry Agent — Autonomous Finance Ops: an event-driven AI agent that turns financial documents and transaction events into completed, auditable finance workflows.

What it does

Cherry Agent operates in the background rather than waiting for a user to continually prompt a chatbot.

When a bill, receipt or bank transaction arrives, the agent can:

Understand the document
Extract supplier, invoice number, date, subtotal, VAT, total and other relevant financial information.
Validate the transaction
Check required fields, detect possible duplicates and identify missing or inconsistent evidence.
Categorise it
Determine the most appropriate accounting category based on the transaction, supplier and available context.
Reconcile it
Search available bank transactions and identify the most likely matching payment using amount, date, supplier and contextual signals.
Assess confidence and risk
Decide whether the workflow can continue automatically or needs human review.
Request approval when necessary
Ambiguous, unusual or higher-risk actions are routed to a human with the relevant evidence and the agent's recommendation.
Complete the finance workflow
Once resolved, Cherry Agent records the reconciliation and maintains the supporting evidence.
Prepare for month-end
The agent can asynchronously identify unreconciled transactions, missing evidence and outstanding exceptions rather than waiting for someone to manually search for them.

The goal is not to replace financial oversight.

The goal is to make humans deal with exceptions and decisions instead of repetitive administration.

Why it is agentic

Cherry Agent is deliberately designed as more than a conversational finance assistant.

A financial event can initiate a workflow without the user asking a question.

The agent then:

observes → reasons → plans → calls tools → validates → acts → records → escalates if required

Each workflow maintains state so that an operation can pause for an approval, resume later, recover from a failed tool call and avoid accidentally processing the same financial event twice.

This makes Cherry Agent particularly suited to long-running finance operations that happen asynchronously in real businesses.

How we're building it

Cherry Agent uses Gemini 3.5 Flash for multimodal understanding and reasoning, orchestrated using Google Agent Development Kit (ADK).

The system is being designed around specialist capabilities rather than one large prompt:

Document Agent — understands invoices and receipts
Categorisation Agent — proposes accounting treatment
Reconciliation Agent — matches financial documents with bank transactions
Risk & Approval Agent — determines whether human review is required
Month-End Agent — finds unresolved finance operations
Evidence Agent — records what happened and why

The event-driven backend uses Google Cloud infrastructure.

Google Cloud Run hosts the agent services.

Pub/Sub carries asynchronous financial events.

Firestore stores workflow and agent state.

Cloud Storage stores documents and supporting evidence.

Gemini / Vertex AI provides multimodal document understanding and agent reasoning.

This architecture allows an incoming financial event to continue through a multi-step workflow without requiring a user to remain online.

Human-in-the-loop finance

Autonomy in financial software needs boundaries.

Cherry Agent therefore uses confidence and risk thresholds.

Routine, deterministic, high-confidence operations can proceed automatically.

Actions involving uncertainty, unusual activity or greater financial impact can be paused and presented to a human for approval.

Every significant action is recorded so that the final result is not simply:

"The AI categorised this transaction."

Instead, we want the system to provide an evidence trail showing what information was used, which tools were called, what decision was made and whether a human approved an exception.

Data sources

The hackathon prototype is designed to work with:

invoices and receipts in PDF or image format
supplier and transaction metadata
bank transaction feeds or sandbox transaction data
accounting categories and policies
workflow state and human approval decisions

The agent separates AI reasoning from deterministic financial controls so that model output alone does not directly authorise sensitive financial actions.

Challenges we're solving

Reliable financial document understanding

Invoices and receipts arrive in different layouts, formats and levels of quality. Multimodal extraction needs to be combined with validation rather than blindly trusting extracted values.

Reconciliation is probabilistic

A bank transaction and invoice do not always contain identical descriptions. Matching therefore requires a combination of deterministic checks and contextual reasoning.

Autonomy without losing control

The hardest question is not whether an AI agent can take an action — it is deciding when it should.

We are designing explicit escalation boundaries so that autonomy increases productivity without removing accountability.

Asynchronous reliability

Finance workflows may wait minutes, hours or days for another event or human approval. The agent therefore needs persistent state, idempotent operations, retries and an audit trail.

What we're learning

One of the most important lessons from building Cherry Agent is that a useful enterprise agent needs more than a powerful model.

It needs tools, state, permissions, deterministic controls, failure handling and observability.

The LLM provides intelligence, but the surrounding architecture determines whether that intelligence can safely perform useful work.

That is especially important in finance, where an agent needs to explain what it did and where humans must remain in control of higher-risk decisions.

What's next

Our immediate goal is to demonstrate one complete autonomous workflow:

Bill / Receipt → Understand → Categorise → Reconcile → Approve if needed → Audit-ready evidence

From there, the same architecture can expand into month-end close, accounts payable, bookkeeping exception handling and other repetitive finance operations.

Cherry Agent's vision is simple: